1. No category

How to render Neural Fields more realistic
Axel Hutt and Meysam Hashemi and Peter beim Graben
hal-01007681, version 1 - 17 Jun 2014
1 Introduction
Complex systems are omnipresent in nature. They exhibit multiple spatial and
temporal scales whose understanding and control is one of the most challenging
problems of research for the last centuries. The notion of complexity is not welldefined [1], but typically a system is said to be complex if, in addition to the multiple scales, there exist self-organised sub-units in the system which are generated
by smaller units. In many biological systems, climbing this hierarchy from smaller
units to larger units implies increasing the spatial and temporal scale.
Before discussing neural systems, we would like to clarify a major scientific
problem in this context: the choice of the description level. Let us assume the task
to describe a water wave mathematically, i.e., derive a mathematical model whose
solution describes the spatiotemporal phenomenon of a wave. Some researchers may
ask: we want to understand how water works, and especially, how a wave is generated. Hence, a straightforward simple and fundamental approach is to take a closer
look at the building blocks of water (remark : this is very similar to the way how
biologists work in neuroscience today). The researchers find H2 O-molecules, study
the properties of the atoms and the inter-molecule interactions like hydrogen bridges
and van-der-Waals bounds. Now, to describe a water wave, one has to study millions
and millions of these molecules and their interactions. To this end, the researchers
Axel Hutt
INRIA CR Nancy - Grand Est, Team Neurosys, Villers-les-Nancy, France, e-mail:
[email protected]
Meysam Hashemi
INRIA CR Nancy - Grand Est, Team Neurosys, Villers-les-Nancy, France, e-mail:
[email protected]
Peter beim Graben
Dept. of German Studies and Linguistics, Humboldt-Universit¨at zu Berlin and Bernstein Center
for Computational Neuroscience, Berlin, Germany e-mail: [email protected]
1
hal-01007681, version 1 - 17 Jun 2014
2
Axel Hutt and Meysam Hashemi and Peter beim Graben
take the reasonable approach to start simulating two, three, then and even hundreds
of these molecules. The numerical analysis is demanding, but still there is no description of the water waves, since the system appears to be so complex taking into
account all the different interactions of molecules and the research in this hard task
got stuck in some way.
Of course it is well-known that it is not necessary to study single molecules to
describe water waves, today we take the famous Navier-Stokes equation (NSE)
which includes the solution for this phenomenon. This equation involves mean fluid
properties, e.g., inner friction and viscosity [2]. Hence the NSE considers average
properties of interacting single molecules. In other words, it does not know single molecules, but provides a powerful description of a large ensemble or mass of
molecules. It allows to describe several different rather complex fluid phenomena,
but of course no phenomena related to single molecules. Hence the NSE equation
provides a very good mathematical description of the system at a macroscopic description level only. Going back to the intended description of water waves by a
single molecule study, this approach is not reasonable, probably it will not lead to
a good description level of macroscopic phenomena and hence it is not constructive.
In todays’ neuroscience, the approach of linking single neuron activity to macroscopic phenomena is attractive, e.g. in the context of cognition [3–5], sleep [6] or
anaesthesia [7,8]. These studies state a relationship between the single neuron activity (microscopic scale) and behavioral phenomenon (macroscopic scale). However,
to our best knowledge it is not understood how the different experimental findings
on different scales are linked to each other, i.e., the link between the two scales is
not understood and no model linking the scales has been developed yet. This situation resembles closely the water wave task described above: it is clear that there is a
relation between small sub-units (molecules or neurons) to large complex systems
units (water wave or cognition), since the dynamics of the large units is generated
by the sub-units, but a link appears to be too complex. Consequently, learning from
physics and the NSE, it is necessary to consider more abstract, intermediate models
whose elements are based on small sub-units properties but which allow to model
large unit phenomena. In other words, it is much more effective to consider mesoscopic population models involving average properties of interacting neurons and
which allow to describe macroscopic experimental phenomena, such as Local Field
Potentials (LFP), encephalographic activity (EEG/MEG) or even behaviour. Promising candidates for such models are neural mass or neural field models. The present
book chapter discusses recent advances in these models rendering the standard neural population models more realistic.
The subsequent sections do not give a complete overview over the recent advances in the field, a recent excellent review article already provides this information [9]. The present chapter first introduces briefly to two types of neural field
models. Then it gives some details of few selected extensions, always introducing
the neuroscientific problem by experimental data before presenting a mathematical
description of the phenomena described.
How to render Neural Fields more realistic
3
2 Two classes of neural field models
2.1 Amari model
The first neural field models developed by Wilson and Cowan [10] and Amari [11]
are continuum limits of large-scale neural networks. Typically, their dynamic variables describe either mean voltage [11] or mean firing rate [10, 12] of a population
element of neural tissue, see also the excellent review article of Bressloff [9]. In
some subsequent sections we consider the paradigmatic Amari equation [11] describing the spatiotemporal dynamics of mean potential V (x,t) over a cortical ddimensional manifold Ω ⊂ Rd :
hal-01007681, version 1 - 17 Jun 2014
τ
∂V (x,t)
= −V (x,t) +
∂t
Z
K(x, y)S[V (y,t)] dy + I(x,t) .
(1)
Ω
with the spatial synaptic kernel K(x, y) which defines the connectivity between site
y ∈ Ω and site x ∈ Ω . The transfer function S is nonlinear and typically of sigmoidal
shape. This model considers external inputs I(x,t), e.g., originating from other extracortical populations and from external stimulation. The model (1) takes into account
a single synaptic time scale τ assuming an exponential synaptic response function.
However we point out that re-scaling of time allows to set τ = 1.
In general, the connectivity kernel K(x, y) fully depends on both sites x and y
reflecting spatial heterogeneity. If the connectivity solely depends on the difference
between x and y, i.e. K(x, y) = K(x − y), then the neural field activity does not depend on specific spatial locations and hence is translational invariant. This case is
called spatially homogeneity [11]. If the connectivity even depends on the distance
between x and y only, i.e. K(x, y) = K(||x − y||), with ||x|| as some norm in Ω , then
the neural field is spatially homogeneous and isotropic [13].
Several extensions of the Amari model (1) are possible, such as the consideration
of finite axonal transmission speeds [14, 15], constant feedback delays [16, 17](see
also section 3), heterogeneity [18, 19] (see also section 4), spike-frequency adaption [20], statistical properties of single neurons [21], the combination of several
brain areas [22], electromagnetic fields [23, 24] and many more [9].
Mathematically, Eq. (1) is an integro-differential equation. Spatially homogeneous (respectively isotropic) neural fields have been intensively studied in the literature due to their nice analytical properties [14, 15, 25, 26]. Moreover, these models
may be transformed to derive partial differential wave equations [25, 27, 28] for certain classes of synaptic kernels.
2.2 Robinson model
As mentioned at the end of the previous paragraph, under certain conditions integrodifferential equations may be transformed to partial differential equations. In the
4
Axel Hutt and Meysam Hashemi and Peter beim Graben
1990’s James Wright and David Liley started developing partial differential equation models for neural population activity [29]. Their work has inspired other teams,
e.g., Peter Robinson and colleagues, who developed a similar neural model which
has been proven to be very successful. This type of neural field model [30–32] is
based on a population-level model of a single thalamo-cortical module consisting of
excitatory (E) and inhibitory (I) cortical population, thalamic relay neurons (S), and
thalamic reticular neurons (R). The average soma membrane potential is modeled
by
¯ ⊗ Ka,b φb (t − τa,b ), a = E, I, S, R
(2)
Va (t) = ∑ h(t)
hal-01007681, version 1 - 17 Jun 2014
b=E,I,R,S
where ⊗ denotes temporal convolution and φb is the pulse firing rate of the population b. The constants Ka,b are the strengths of the connections from population
of type b to population of type a. The delay term, τa,b is zero for intra-cortical and
intra-thalamic connections and non-zero for thalamocortical or corticothalamic connections [31].
The model assumes that only axons of excitatory cortical neurons are long
enough to emit axonal propagating pulses. Moreover, φE obeys the damped oscillator equation
DφE = S(Va ),
(3)
with the operator D
D=
1 ∂
+1
γ ∂t
2
.
(4)
¯ denotes the mean synaptic response function
In Eq. (2) h(t)
¯ = αβ e−αt − e−βt ,
h(t)
β −α
(5)
where α and β are the synaptic decay and rise rate of synaptic response function,
respectively.
3 Delayed nonlocal feedback between populations
In neural fields, one might include delays in several ways. The finite axonal transmission delay is proportional to the fraction of distance between two spatial locations and transmission speed and takes into account the finite propagation speed of
action potentials along axonal branches [33], or in more general terms, it considers
the finite-time interaction between two elements in a spatially extended system [28].
In addition, one could argue that delayed interactions happen on a single-neuron
scale between neurons and it is more reasonable to treat these inter-neuron delays
as a kind of effective delay [17, 34]. This latter type of delay is constant. In addition
to these two delay types, the nonlocal feedback delay takes into account the finite
axonal transmission speed along axonal pathways between two brain areas. Since
How to render Neural Fields more realistic
5
this axonal pathway has a finite defined length, the transmission delay is fixed and
hence also constant [16]. Variations of all these delay types may be considered by
distributed transmission speeds and/or distributed delays [15, 35]. The subsequent
paragraphs consider constant delays reflecting the finite transmission speed along
axonal branches between brain areas.
hal-01007681, version 1 - 17 Jun 2014
3.1 The primary sensory area in weakly-electric fish
To examine the neural decoding in weakly electric fish, Doiron et al. [36] had performed an experimental in vivo stimulation study. A dipole was placed near the
skin of a fish to stimulate only a part of the receptive field. Figure 1 sketches the
experimental setup, the spike autocorrelation A(t) and the interspike interval (ISI)
histograms of a typical pyramidal cell response to a local and global stimulus. The
stimulus was temporal noise evoking spatially weakly correlated sensory receptor
activity for local stimulation and strong spatial correlations in the receptor dynamics
in case of the experimentally global stimulation. The global random stimulus evokes
bursting in the neuron activity whereas the local stimulus evokes a single principal
firing mode. This experiment raises the question how the spatial correlations in the
input stimulus interact with those imposed by the physiological system.
Fig. 1 Experimental setup and firing statistics of stimulation experiment [36]. In (a), the electric
skin stimulation is local (left panel) and induces a single main oscillation as seen in the spike
autocorrelation function A(τ) (center panel) and the corresponding histogram (right panel). In (b)
the stimulus was global inducing an additional oscillation mode. Taken from [36] by permission.
A rather simple population model considers the primary sensory areas in the
electro-sensory system of weakly electric fish [37], but similar configurations can
also be found in parts of the vertebrate brain [38]. The model sketched in Fig. 2 [39]
6
Axel Hutt and Meysam Hashemi and Peter beim Graben
hal-01007681, version 1 - 17 Jun 2014
is made up of the ELL, a layer of pyramidal cells driven by the primary receptors
that receive an external stimulus, and the higher area N p. These areas are spatially
coupled via a delayed topographic feedback with connectivity kernels Ken (x) and
Kne (x) which reflect connections from the N p(n) to the ELL(e) and vice versa,
respectively, see Fig. 2. The neurons in both populations have insignificant direct
couplings and the coupling from the ELL to N p is excitatory and delayed in time by
τ1 , the coupling back to the ELL is inhibitory with delay τ2 . Moreover, according
to the experimental setup in [36], the model considers excitatory spatiotemporal
stimuli I(x,t) to the ELL.
Fig. 2 Topography of the delayed feedback model. The plus and minus signs indicate the excitatory
and inhibitory connections, respectively.
The aim of the study is to learn more about the mechanism how to change the
principal oscillation frequency by the properties of an external stimulus. The model
considers spatiotemporal noise with a well-defined and adjustable spatial correlation
length. The population model [39] describes two coupled neural fields whose activities are strongly related to experimentally observable local field potentials [40].
Then the theoretical power spectrum in the ELL reads
Z ∞
P(ν) =
˜
R(ν, l)C(l)dl
(6)
−∞
˜ is the scaled Fourier transwhere R(ν, l) is the spectral response function and C(l)
form of the input correlation function. It turns out, that the power spectrum (6) does
not depend on the spatial scale of the feedback loop σ f and the input correlation
scale σi independently, but just depends on their ratio, called η. This finding reflects
the coupling of the spatial scale of the external input to the intrinsic spatial scale of
the system.
Figure 3 shows the resulting power spectra for two values of the spatial scale
ratio η. We observe that a small ratio η = σ f /σi 1 generates a spectral peak at
about 20Hz, whereas the large ratio η 1 generates a power peak at about 0Hz.
Hence retaining the topographic feedback but decreasing the input correlation function from large values of σi = σ f /η (global noise) to small values of σi (local noise)
switches the spectral peak, similar as observed experimentally by Doiron et al. [36].
How to render Neural Fields more realistic
7
hal-01007681, version 1 - 17 Jun 2014
Fig. 3 Theoretical power spectrum computed for η = 40 (solid line) and η = 1/40 (dashed line).
˜
Fig. 4 The response function R(ν, l) and the integrand of the power spectrum integral R(ν, l)C(l)
for (a,b) σi = 40σ f (global noise), (c,d) σi = σ f /40 (local noise). Taken from [39] by permission.
To understand this, Figure 4 shows the response function R(ν, l) and the inte˜ in the definition of the power spectrum (6). The response function
grand R(νl )C(l)
˜
R and RC have a single maximum at about 20Hz for global noise (Fig. 4(a,b)),
whereas R and RC˜ have two local maxima at 20Hz and 0Hz (Fig. 4(c,d)). Since the
peak of RC˜ at 0Hz is broader than the peak about 20Hz and the power spectrum is
˜ cf. Eq. (6), the contribution of RC˜ to the power at 0Hz exceeds
the integral over R(l),
the contribution at 20Hz yielding a strong peak at 0Hz.
8
Axel Hutt and Meysam Hashemi and Peter beim Graben
This result reveals that the spectral peak in Fig. 3 at 0Hz results from the selection
of one mode out of two possible modes at frequencies of 0Hz and 20Hz, whereas the
spectral peak at 20Hz is the only oscillation mode present in the system. The switch
between these two configurations depends on the spatial correlation of the stimulus
noise. These modes reflect activity subnetworks from which only one is engaged.
hal-01007681, version 1 - 17 Jun 2014
3.2 General anaesthesia
General anaesthesia is an important medical application in today’s hospital surgery,
but its underlying neural interactions is still a mystery. In the last decades, general
anaesthesia has attracted theoretical researchers [41–45]. Most theoretical studies
aim to explain signal features of electroencephalographic data (EEG) observed during anaesthesia, such as the attenuation or enhancement of α−activity accompanied
by a subsequent enhancement of δ −activity while increasing anaesthetic concentration [46, 47], cf. Fig. 5. The subsequent paragraphs show how neural field models
Fig. 5 The power spectra measured in frontal EEG electrodes in the absence of anaesthesia (blue
line) and during propofol anaesthesia (red line) in a group of subjects (a) and for a single subject
(b). Taken from [30] by permission.
may explain on the power enhancement and the frequency shift of maximum power
while increasing the anaesthetic concentration.
To this end, we consider a derivative of the Robinson model introduced in section 2.2 and introduce a new sigmoid function derived from properties of type-I
neurons [48]
S(Va ) = F(Va , 0) − F(Va , γ),
(7)
with
Smax
F(Va , γ) =
2
2 2
Va − θ − γσ 2
√
1 + erf
e−γ(Va −θ )+γ σ /2 ,
2σ
(8)
How to render Neural Fields more realistic
9
in which the parameter γ < ∞ takes into account the properties of type I-neurons,
Smax is the maximum population firing rate, θ is the mean firing threshold, and σ is
related to the standard deviation of firing thresholds in the populations.
External input to the system can be considered as a non-specific input to relay
neurons as
√
(9)
φN = hφN i + 2κξ (t),
hal-01007681, version 1 - 17 Jun 2014
where hφN i is its mean value and ξ (t) is Gaussian white noise of strength κ with
zero mean.
It is well-known that general anaesthetics (GA) bind directly to sensitive target
receptors. For instance, a large number of studies link the effect of many GAs to
altered function of the GABAA receptors. Recent clinical findings have revealed
several sites of elicited anaesthetic action of the anaesthetic propofol in the human
brain, e.g., propofol suppresses field potentials in the rat thalamus and cortex [49].
In detail, the anaesthetic propofol increases the decay time constant of synaptic
GABAA receptors, and hence increases the total charge transfer in these synapses
but not that of excitatory synapses [50]. To integrate physiological observations into
neural models such as the Robinson model detailed in section 2.2, the anaesthetic action on synaptic receptors is modelled by α → α/p with p ≥ 1 leading to a decrease
of the decay rate constant α, or equivalently, an increase in decay time constant of
GABAA receptors [51]. The factor p = 1 reflects absent anaesthetic action, i.e, the
baseline condition. The model considers inhibitory synapses at excitatory neurons
(factor p1 ), and at inhibitory neurons (factor p2 ) and at thalamic relay neurons (factor p3 ).
To compute the power spectrum of the system, we consider the stationary state
of Eq. (2), which obeys dVa (t)/dt = 0 where Va is taken from Eq. (2). Increasing
the anaesthetic concentration, i.e. increasing the three factors p1 , p2 and p3 changes
the stationary states dependent on the relation of these three factors, cf. Fig. 6. In (a)
the two lower stationary states collide to a single state whereas in (b) the two upper
states collide. This difference indicates two fundamentally different mechanisms
which may yield different power spectra.
The power spectrum characterises small fluctuations about this stationary state.
Assuming that excitatory activity generates the EEG, and by virtue of the specific
choice of external input to relay neurons, the power spectrum of the EEG depends
just on one matrix component of the Greens function [51] by
√ 2
(10)
PE (ω) = 2κ 2π G˜ 1,3 (ω) .
in which G˜ 1,3 (ω) is a matrix element of the 4 × 4 matrix
hal-01007681, version 1 - 17 Jun 2014
10
Axel Hutt and Meysam Hashemi and Peter beim Graben
Fig. 6 The stationary states and the nonlinear gain dS/dV computed at the lowest stationary state
of pyramidal neurons VE subjected to the factor p2 . (a), (b) p1 = p3 = 1 + 0.3(p2 − 1), (c), (d)
p1 = p2 = p3 . We observe three states in (a) and (c) for p2 = 1 where (a) the two lower states
collide and (c) the two upper states collide. The center branch (dashed red) is linearly unstable,
whereas the other branches are linearly stable. For clarification, the lower branches are shown in
the insets. Parameters are Smax = 250Hz, Vth = 15mV, γ = 0.08mV and σ = 10mV.
−1
˜
L˜ KD11 − K1 −K2 −K3 e−iωτ 0
1  −K4 L˜ − K5 −K6
0 

˜
G(ω)
=√ 
−iωτ

˜
0
L
−K8 
2π −K7 e
−K9 e−iωτ
0
−K10
L˜

(11)
with L˜ = (1 + iω/α)(1 + iω/β ), D˜ = (1 + iω/γ)2 , and the constants Ki , i =
1, 2, ..., 11 are proportional to the synaptic strengths and the nonlinear gains ∂ S/∂V
computed at the stationary state of the system.
Figure 7 shows the theoretical power spectrum PE in the baseline condition and
after the administration of propofol for two different relations of the anaesthetic
factors p1 , p2 and p3 . At first we note that the spectra resemble well the spectrum obtained from experimental observations: increasing p2 in a specific relation
to the p1 and p3 yields increases in delta and theta power as well as more pronounced alpha oscillation with increased peak-frequency. The dynamical analysis
of the model [30] reveals three resonances in the baseline condition, including an
oscillatory resonance corresponding to the peak in the alpha-band and a pair of
zero-frequency resonances. Increasing the anaesthetic concentration diminishes the
How to render Neural Fields more realistic
11
hal-01007681, version 1 - 17 Jun 2014
damping rate of alpha resonances (and hence increases its magnitude) while its frequency increases. Hence increases in alpha power and its peak-frequency results
from the approach of the system of an oscillatory instability [52]. Moreover, the two
zero-frequency resonances collide and gradually increases in frequency leading to a
magnitude increases in delta and theta power (solid red line in Fig. 7.)
Fig. 7 Theoretical power spectrum in the baseline condition (p1 = p2 = p3 = 1.0 encoded in blue)
and in the anesthesia condition (red line). (a) p1 = p3 = 1 + 0.3(p2 − 1) and p2 = 1.15 (red line)
and (b) p1 = p2 = p3 = 1.15 (red line).
Summarizing, previous studies [30, 51] have revealed the differential role of
synaptic inhibition at GABAergic receptors located on the dendrites of different neural types. Moreover, synaptic potentiation within local cortical inhibitory neurons
suffices to reproduce experimental observation in EEG during anesthesia [30, 52].
4 Heterogeneous neural fields
Previous experimental studies on the structure of biological neural networks and
their connections, e.g., as the exciting work of Hellwig [53], reveals a large rather
simple mesoscopic structure at the spatial scale of several hundreds of micrometers
with an underlying more complex structure at smaller scales. For instance, [53] reveals a rather homogeneuous Gaussian or exponential connectivity distribution of
single neurons in layers II and III in rat visual cortex on a scale of ∼ 500µm with an
overlying complex connectivity structure at much shorter scales reflecting heterogeneous connections. Hence, a homogeneous model serves as a first approximation
12
Axel Hutt and Meysam Hashemi and Peter beim Graben
at larger spatial scales, whereas heterogeneous connections need to be considered in
more realistic models.
Several previous theoretical studies have investigated the effect of heterogeneous
connections in neural fields, primarily in the context of pattern formation [54–58].
Most of the previous studies consider the heterogeneous fields as small heterogeneous perturbations of homogeneous states. The subsequent paragraphs show a new
approach taking into account the heterogeneous structure not as a limit of the homogeneous case.
Starting from the Amari equation (1), we expand the integral into a power series
[19]
∂V (x,t)
+V (x,t)
Z∂t
=
(12)
Z Z
K1 (x, y)V (y,t) dy +
Ω
K2 (x, y, z)V (y,t)V (z,t) dy dz ,
Ω
Ω
hal-01007681, version 1 - 17 Jun 2014
where the kernels K1 and K2 have to be reconstructed from experimental data.
Event-related potentials
In order to illustrate this, we combine a recently developed technique for nonlinear data analysis [18] with our new method for training heterogeneous neural
fields [19] [19]. Figure 8 shows the grand average event-related potentials (ERP)
Un (t), n = 1 . . . , N of N = 25 recording electrodes from a language processing experiment of subject-object ambiguities in German [59]. The first column displays
the control condition with a canonical subject-verb-object order: “Die Rednerin hat
den Berater beim Kongress gesucht” (“The speaker has sought the advisor at the
congress”), while the second column (Fig. 8(e –h)) shows ERPs for a non-canonical,
yet in German grammatical, object-verb-subject order: “Die Rednerin hat der Berater beim Kongress gesucht” (“The speaker has been sought by the advisor at the
congress”). The panels (a) and (b) in the first row of Fig. 8 present the ERPs averaged over 14 subjects elicited by the disambiguating article “den” vs. “der” at time
zero. Each trace shows the voltage of one of N = 25 EEG electrodes. Comparing
panels (a) and (b) in Fig. 8, one recognises a difference between conditions at about
600ms after onset of the critical stimulus, known as the P600 ERP component.
Heteroclinic orbits
To analyze these spatiotemporal ERP patterns we apply our recently developed recurrence domain segmentation technique [18] whose results are shown in Fig. 8(c)
and (d). This method computes the recurrence plots from the N-dimensional trajectories of each condition using a ball-size ε > 0 as parameter. Here, we optimized ε
hal-01007681, version 1 - 17 Jun 2014
How to render Neural Fields more realistic
13
Fig. 8 Experimental event-related potentials (ERP) [59] in two experimental conditions (left and
right column) and their interpretation as heteroclinic orbits. (a,b) ERPs of N = 25 single EEG
time series, (c,d) temporal sequence of extracted recurrence domains, (e,f) the spatial EEG activity
distributions on the scalp (ERP components, seen from above, nose on top, back of head on bottom of the maps) averaged over the time windows of recurrence domains shown in panels (c,d),
projections of the multivariate EEG signal (shown in panels (a,b)) on the corresponding ERP component maps (shown in panels (e,f), (g,h) sequence of heteroclinic saddles (SHS) modeled by the
Lotka-Volterra model.
according to a reasonable heuristics for obtaining a relatively low number of recurrence domains with relatively long dwell times and good (visual) discrimination of
conditions which worked quite well for ε = 1.9µV . Following [18], the recurrence
plots are interpreted as rewriting grammars to replace large time indices by lower,
recurrent, ones. These transformed time indices are plotted in Fig. 8(c) and (d) as
different colors. The processing differences between conditions are clearly seen as
different segments, indicating that ERP components, such as the P600, can be regarded as recurrence domains, or, in first approximation, as saddle nodes [60, 61].
These saddles are estimated in Fig. 8(c) and (g) as follows: For each segment k of
Fig. 8(c) and (d) we compute its center of gravity, i.e. its temporal average, to obtain
a spatial EEG activity distribution on the scalp, namely a spatial voltage distribution
Uk (z), k = 1 . . . , P with number of segments P and where z is the spatial location on
the scalp. The patterns Uk (z) are activity patterns extrapolated in space on the basis
of the average ERP components which are discrete in space. They are shown in Fig.
8(e) and (f).
These distributions are assumed to originate from underlying neural patterns embed-
14
Axel Hutt and Meysam Hashemi and Peter beim Graben
ded in neural populations, e.g. sub-networks in neural populations. These underlying corresponding neural patterns may be called Vk (x). Let us call V (x,t) the neural
population activity generating the EEG signal. Then the Moore-Penrose pseudoinverses Vk (x)+ of Vk (x) form a bi-orthogonal basis with the patterns Vk (x), such that
the projections
Z
Vk (x)+V (x,t)dx
ξk (t) =
Ω
serve as order parameters in a decomposition
V (x,t) = ∑ ξk (t)Vk (x) .
hal-01007681, version 1 - 17 Jun 2014
k
The ξk (t) are plotted in Figs. 8(g) and (h) where each trace depicts the projection
onto one segment from the recurrence domain analysis in Figs. 8(c) and (d).
The dynamics of the order parameters {ξk } is of high temporal complexity, indicating that a large number of time scales is involved in the ERP language processing dynamics. However, in a first approximation we assume that the recurrence
domains are in fact isolated saddles that are connected along stable heteroclinic sequences (SHS). This can be modeled through winnerless competition in a network
of Lotka-Volterra populations [62,63]. This assumption leads to the order parameter
dynamics shown in 8(g,h). From the growth rates σk > 0, and interaction weights
ρk j > 0, ρkk = 1, of the Lotka-Volterra dynamics and from the bi-orthogonal pattern
systems V j (x),Vk (x)+ , one can construct the kernels of the neural field equation (12)
as
K1 (x, y) = ∑(σk + 1)Vk+ (y)Vk (x)
k
K2 (x, y, z) = ∑ ρk j σ jVk+ (y)V j+ (z)Vk (x) ,
kj
We observe that the kernel K1 (x, y) describes Hebbian synapses between sites
y and x trained with pattern sequence Vk [19]. Interestingly, this memory storage
mechanism resembles well the storage of patterns in a Bidirectional Associative
Memory (BAM) [64].
Moreover, to our best knowledge the three-point kernel K2 (x, y, z) has not been studied yet in the context of neural fields. It further generalizes Hebbian learning to
interactions between three sites x, y, z ∈ Ω . Interestingly, one can write
K1 (x, y) = ∑(σk + 1)K1,k (x, y)
k
"
K2 (x, y, z) = ∑
j
#
∑ ρk j K1,k (x, y)
σ jV j+ (z) .
k
with K1,k (x, y) = Vk+ (y)Vk (x). According to Eq. (13) the three-point kernel can be see
as a linear superposition of two-point kernels weighted by the interaction weights
ρk j and the underlying stored patterns V j+ (z) and its corresponding growth rates σk .
How to render Neural Fields more realistic
15
The additional dependence of the spatial location z is reasonable in heterogeneous
fields: the action at spatial location x does not only depend on the activity in spatial location y (as assumed in homogeneous systems), but may also depend on the
path between x and y which in turn depends on the underlying memory structure.
This argument motivates the presence of an additional spatial variable z, but does
neither explain the terms in K2 (x, y, z) nor its dependence on the specific choice of
the heteroclinic dynamics (here a Lotka-Volterra system). Future work will attempt
to elaborate more details on the presence of three-point kernels.
hal-01007681, version 1 - 17 Jun 2014
5 Conclusion
This chapter has presented some theoretical neural field studies describing the power
spectra in homogeneous neural fields and the nonlinear dynamics in heterogeneous
neural fields. In this context, we want to mention recent extensions of standard neural field models by incorporating single neuron properties [48, 65–67].
Acknowledgements
The authors thank Stefan Frisch and Heiner Drenhaus for conducting the ERP experiment. AH and MH acknowledge funding from the European Research Council for
support under the European Union’s Seventh Framework Programme (FP7/20072013) ERC grant agreement No.257253. PbG acknowledges financial support through
a Heisenberg Fellowship of the German Research Foundation DFG (GR 3711/1-2).
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
P. Grassberger, Int. J. Theor. Phys. 25(9), 907 (1986)
L. Landau, E. Lifshitz, Fluid Mechanics (Butterworth-Heinemann, 1987)
C.G. Gross, Neuroscientist 8(5), 512 (2002)
N.E. Barraclough, D.I. Perrett, Phil. Trans. R. Soc. B 366(1571), 1739 (2011)
R.Q. Quiroga, L. Reddy, G. Kreiman, C. Koch, I. Fried, Nature 435(7045), 1102 (2005)
V.V. Vyazovskiy, K.D. Harris, Nature Rev. Neurosci. 14, 443 (2013)
B. Antkowiak, Anesthesiology 91, 500 (1999)
D. Belelli, N.L. Harrison, J. Maguire, R.L. Macdonald, M.C. Walker, D.W. Cope, J. Neurosc.
29(41), 12757 (2009)
P.C. Bressloff, J. Phys. A 45(3), 033001 (2012)
H. Wilson, J. Cowan, Kybernetik 13, 55 (1973)
S.I. Amari, Biol. Cybern. 27, 77 (1977)
D. Jancke, W. Erlhagen, H.R. Dinse, A.C. Akhavan, M. Giese, A. Steinhage, G. Sch¨oner, J.
Neurosci. 19(20), 9016 (1999)
S. Coombes, N.A. Venkov, L. Shiau, I. Bojak, D.T.J. Liley, C.R. Laing, Phys. Rev. E 76(5),
051901 (2007)
16
Axel Hutt and Meysam Hashemi and Peter beim Graben
14.
15.
16.
17.
18.
A. Hutt, N. Rougier, Phys. Rev. E. 82, R055701 (2010)
A. Hutt, L. Zhang, J. Math. Neurosci. 3, 9 (2013)
A. Hutt, Phys. Rev. E 70, 052902 (2004)
A. Roxin, N. Brunel, D. Hansel, Phys.Rev.Lett. 94, 238103 (2005)
P. beim Graben, A. Hutt, Physical Review Letters 110(15), 154101 (2013). DOI
10.1103/PhysRevLett.110.154101
P. beim Graben, A. Hutt, EPJ Nonlin. Biomed. Phys. accepted (2014)
S. Coombes, M. Owen, Phys. Rev. Lett. 94, 148102 (2005)
O.D. Faugeras, J.D. Touboul, B. Cessac, Front. Comput. Neurosci. 3, 1 (2008)
N. Rougier, J. Vitay, Neural Networks 19(5), 573 (2006)
P. beim Graben, S. Rodrigues, Frontiers in Computational Neuroscience 6(100) (2013). DOI
10.3389/fncom.2012.00100
P. beim Graben, S. Rodrigues, in Neural Fields: Theory and Applications, ed. by S. Coombes,
P. beim Graben, R. Potthast, J.J. Wright (Springer, Berlin, 2014)
S. Coombes, G. Lord, M. Owen, Physica D 178, 219 (2003)
S. Folias, P. Bressloff, SIAM J. Appl. Math 65, 2067 (2005)
V.K. Jirsa, H. Haken, Phys. Rev. Lett. 77(5), 960 (1996)
A. Hutt, Phys. Rev. E 75, 026214 (2007)
J. Wright, D. Liley, Behav. Brains Sc. 19, 285 (1996)
R. Hindriks, M.J.A.M. van Putten, Neuroimage 60, 2323 (2012)
J. Victor, J. Drover, M. Conte, N. Schiff, Proceed. Natl. Acad. Science USA 118, 15631 (2011)
P. Robinson, C.J.Rennie, D.L.Rowe, S.C.O’Connor, Human Brain Mapping 23, 53 (2004)
A. Hutt, M. Bestehorn, T. Wennekers, Network: Comput. Neural Syst. 14, 351 (2003)
A. Roxin, N. Brunel, D. Hansel, Prog. Theor. Phys. 161, 68 (2006)
F.M. Atay, A. Hutt, SIAM J. Appl. Dyn. Syst. 5(4), 670 (2006)
B. Doiron, M. Chacron, L.Maler, A. Longtin, J. Bastian, Nature 421, 539 (2003)
N.J. Berman, L. Maler, J. Exp. Biol. 202, 1243 (1999)
E. Ahissar, D. Kleinfeld, Cerebral Cortex 13, 53 (2003)
A. Hutt, C. Sutherland, A. Longtin, Phys.Rev.E 78, 021911 (2008)
P. Nunez, in Neocortical dynamics and human EEG rhythms, ed. by P. Nunez (Oxford University Press, New York-Oxford, 1995), pp. 475–533
M. Steyn-Ross, D. Steyn-Ross, J.W. Sleigh, D.T.J. Liley, Phys. Rev. E 60(6), 7299 (1999)
I. Bojak, D. Liley, Phys. Rev. E 71, 041902 (2005)
A. Hutt, J. Sleigh, A. Steyn-Ross, M.L. Steyn-Ross, Scholarpedia 8(8), 30485 (2013)
S. Ching, A. Cimenser, P.L. Purdon, E.N. Brown, N.J. Kopell, Proc. Natl. Acad. Sci. USA
107(52), 22665 (2010)
M.M. McCarthy, E.N. Brown, N. Kopell, J. Neurosci. 28(50), 13488 (2008)
A. Cimenser, P.L. Purdon, E.T. Pierce, J.L. Walsh, A.F. Salazar-Gomez, P.G. Harrell,
C. Tavares-Stoeckel, K. Habeeb, E.N. Brown, Proc. Natl. Acad. Sci. USA 108(21), 8832
(2011)
M. Murphy, M.A. Bruno, B.A. Riedner, P. Boveroux, Q. Noirhomme, E.C. Landsness, J.F.
Brichant, C. Phillips, M. Massimini, S. Laureys, G. Tononi, M. Boly, Sleep 34(3), 283 (2011)
A. Hutt, L. Buhry, submitted (2013)
B. Antkowiak, Brit. J. Anaesth. 89(1), 102 (2002)
A. Kitamura, W. Marszalec, J. Yeh, T. Narahashi, J. Pharmacol. 304(1), 162 (2002)
M. Hashemi, A. Hutt, submitted (2013)
A. Hutt, Front. Comp. Neurosci. 7, 2 (2013)
B. Hellwig, Biol. Cybernetics 82, 11 (2000)
V. Jirsa, J. Kelso, Phys.Rev.E 62(6), 8462 (2000)
Z.P. Kilpatrick, S.E. Folias, P.C. Bressloff, SIAM J. Applied Dynanmical Systems 7(1), 161
(2008)
S. Coombes, C. Laing, H. Schmidt, N. Svanstedt, J. Wyller, Discrete Contin. Dyn. Syst. A 32,
2951 (2012)
H. Schmidt, A. Hutt, L. Schimansky-Geier, Physica D 238(14), 1101 (2009)
P.C. Bressloff, Physica D 155, 83 (2001)
19.
20.
21.
22.
23.
hal-01007681, version 1 - 17 Jun 2014
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
How to render Neural Fields more realistic
59.
60.
61.
62.
63.
hal-01007681, version 1 - 17 Jun 2014
64.
65.
66.
67.
17
P. beim Graben, S. Gerth, S. Vasishth, Cognitive Neurodynamics 2(3), 229 (2008)
A. Hutt, H. Riedel, Physica D 177(1-4), 203 (2003)
A. Hutt, International Journal of Bifurcation and Chaos 14(2), 653 (2004)
V.S. Afraimovich, V.P. Zhigulin, M.I. Rabinovich, Chaos 14(4), 1123 (2004)
M.I. Rabinovich, R. Huerta, P. Varona, V.S. Afraimovichs, PLoS Computational Biolog 4(5),
e1000072 (2008)
B. Kosko, IEEE Transactions on Systems, Man and Cybernetics 18(1), 49 (1988)
A. Hutt, Cogn. Neurodyn. 6, 227 (2012)
J. Baladron, D. Fasoli, O. Faugeras, J. Touboul, J. Math. Neurosci. 2, 10 (2012)
P. Bressloff, SIAM J. Appl. Math 70, 1488 (2009)